Swapped back to qwen3-tts

docker-compose.yml

@@ -27,7 +27,8 @@ services:
       --mem-fraction-static 0.8
       --max-running-requests 128
       --chunked-prefill-size 4096
-      --context-length 32768
+      --context-length 65536
+      --kv-cache-dtype fp8_e4m3
       --trust-remote-code
       --enable-piecewise-cuda-graph
       --schedule-policy lpm
@@ -40,20 +41,21 @@ services:
               device_ids: ['0', '7']
               capabilities: [gpu]
 
-  # --- TTS ---
-  # Physical: 7 | Container: 0
+# --- TTS ---
   tts:
-    build: ./swarm-control/indra-tts-server
+    build:
+      context: .  # This allows the build to see the Qwen3-TTS folder at the root
+      dockerfile: ./swarm-control/indra-tts-server/Dockerfile
     image: swarm-tts
     depends_on:
       - persona
     environment:
       - CUDA_VISIBLE_DEVICES=0
-      - PYTHONPATH=/app:/app/Qwen3-TTS
+      - PYTHONPATH=/app:/app/Qwen3-TTS # Keep this so the app finds the local code
       - NVIDIA_DRIVER_CAPABILITIES=all
     volumes:
       - /mnt/nvme3n1/swarm/voice-samples:/mnt/nvme3n1/swarm/voice-samples:ro
-      - /mnt/nvme3n1/swarm/Qwen3-TTS:/app/Qwen3-TTS
+      - /mnt/nvme3n1/swarm/Qwen3-TTS:/app/Qwen3-TTS # Keep this for live code edits
     ports:
       - "8002:8002"
     deploy:
@@ -109,7 +111,7 @@ services:
       --host 0.0.0.0
       --hf-chat-template-name tool_use
       --mem-fraction-static 0.95
-      --context-length 32768
+      --context-length 131072
       --trust-remote-code
       --tool-call-parser qwen3_coder
     deploy:
@@ -180,8 +182,8 @@ services:
       --tp 2
       --port 3003
       --host 0.0.0.0
-      --mem-fraction-static 0.85
-      --context-length 65536     
+      --mem-fraction-static 0.80
+      --context-length 131072     
       --kv-cache-dtype fp8_e4m3  
       --trust-remote-code
       --tool-call-parser gemma4
@@ -206,7 +208,7 @@ services:
     ports:
       - "8000:8000"
     command: >
-      --model-id jinaai/jina-embeddings-v2-base-code
+      --model-id google/embeddinggemma-300m
       --max-client-batch-size 1024
     deploy:
       resources: